If you're in the Claude ecosystem, Artifacts can fill a similar purpose. They've gotten really good, if you haven't checked back on them in a while. The recent Docs and Slides update especially is fantastic for working on deliverables in those formats.
Watching for changes is the agent's job by design. Pinrail doesn't know about your repo. It's a generic review surface for whatever the agent sends, whether that's a diff, an email or an image. The agent knows when its input has changed, so it's the one that resubmits or withdraws the previous submissions. And you can instruct your agent to do so with a prompt, or if you wish do it automatically with a custom script
very nice. ive been hopping around threads and browser windows trying to find the right PLACES to unblock the ai, its useful to have it in a single app
Agents can now do real work on their own: they review pull requests, draft emails, write documents and generate images. The hard part is no longer getting the work done, but reviewing it well before it is used.
Pinrail is a desktop app for that review. The agent submits what it made and waits. You see it in a view made for that kind of content, such as comments on a diff, a set of draft emails or an image you can mark up, and you decide item by item. The agent then continues with your decision.
I built it because I kept approving agent work in the worst possible place: a terminal scrollback. When the agent drafted twelve review comments for a pull request, I either let it post all of them or argued about them one by one in chat. Neither worked well, so I had the agent write its comments into a custom HTML page that showed them on the diff. There I could accept some, reject others with a reason, edit one, and hand the whole decision back at once. That page worked so well that I wanted the same for everything else my agents asked me about, and that is how Pinrail started.
It works like this. In the agent's instructions you name the steps that need a person, for example "before posting review comments, submit them to Pinrail and wait". When the agent reaches that step, it runs:
The review appears in the app's inbox and you get a notification. You decide in a view made for that kind of content. The command then prints your decision as markdown, or as JSON with `--json`, and exits with a code that says how the review ended. If you ask for changes, the agent sends a new round with `--revises`, and the app shows your previous verdicts next to it. If you discard a review, the command exits 5 and the agent stops the work the review was gating.
Some details:
- It works with any agent that can run a command: Claude Code, Codex, Cursor, OpenCode, Gemini CLI, or a CI job. The app can install a skill that teaches the agent how to submit and wait.
- Each kind of review is a plugin: a manifest, two JSON Schemas (what the agent sends and what comes back) and an HTML view, which runs in a sandboxed iframe. The app comes with plugins for lists, questions, code review, images and markdown documents. Others cover emails, calendars, HTML pages, logos and 3D models. `pinrail plugins new` creates your own.
- Everything is local. The app runs its server on 127.0.0.1, stores reviews in SQLite, and has no account and no telemetry.
- It is built with a Rust core, Tauri and React. It runs on macOS and Linux. Windows is planned.
- It is free and open source under Apache 2.0.
[sorry, obligatory rotten emacs user who has to turn everything into a minor mode or at least configure emacs-like keyboard nav everywhere on my workstation]
Thanks for this.
I did some cursory poking around, and I'm glad to say, pinrail does not appear to be completely hopeless.
Obviously, (a lot) more work is needed for even a starter emacs-flavored keymap or an emacs-native package to integrate with something like org-agenda.
But again, I'm glad that it does not (initially) appear to me that pinrail has painted itself into a corner in a way that would prevent that work from coming to fruition. And OP even seems amenable to FOSS contributions.
If you're in the Claude ecosystem, Artifacts can fill a similar purpose. They've gotten really good, if you haven't checked back on them in a while. The recent Docs and Slides update especially is fantastic for working on deliverables in those formats.
This looks great, can this be a claude mod ?
The exitcode contract is such a clean idea
Is there a way to mark an approval as stale if the diff changes while you're reviewing it?
Watching for changes is the agent's job by design. Pinrail doesn't know about your repo. It's a generic review surface for whatever the agent sends, whether that's a diff, an email or an image. The agent knows when its input has changed, so it's the one that resubmits or withdraws the previous submissions. And you can instruct your agent to do so with a prompt, or if you wish do it automatically with a custom script
very nice. ive been hopping around threads and browser windows trying to find the right PLACES to unblock the ai, its useful to have it in a single app
Hi HN, author here.
Agents can now do real work on their own: they review pull requests, draft emails, write documents and generate images. The hard part is no longer getting the work done, but reviewing it well before it is used.
Pinrail is a desktop app for that review. The agent submits what it made and waits. You see it in a view made for that kind of content, such as comments on a diff, a set of draft emails or an image you can mark up, and you decide item by item. The agent then continues with your decision.
I built it because I kept approving agent work in the worst possible place: a terminal scrollback. When the agent drafted twelve review comments for a pull request, I either let it post all of them or argued about them one by one in chat. Neither worked well, so I had the agent write its comments into a custom HTML page that showed them on the diff. There I could accept some, reject others with a reason, edit one, and hand the whole decision back at once. That page worked so well that I wanted the same for everything else my agents asked me about, and that is how Pinrail started.
It works like this. In the agent's instructions you name the steps that need a person, for example "before posting review comments, submit them to Pinrail and wait". When the agent reaches that step, it runs:
The review appears in the app's inbox and you get a notification. You decide in a view made for that kind of content. The command then prints your decision as markdown, or as JSON with `--json`, and exits with a code that says how the review ended. If you ask for changes, the agent sends a new round with `--revises`, and the app shows your previous verdicts next to it. If you discard a review, the command exits 5 and the agent stops the work the review was gating.Some details:
- It works with any agent that can run a command: Claude Code, Codex, Cursor, OpenCode, Gemini CLI, or a CI job. The app can install a skill that teaches the agent how to submit and wait. - Each kind of review is a plugin: a manifest, two JSON Schemas (what the agent sends and what comes back) and an HTML view, which runs in a sandboxed iframe. The app comes with plugins for lists, questions, code review, images and markdown documents. Others cover emails, calendars, HTML pages, logos and 3D models. `pinrail plugins new` creates your own. - Everything is local. The app runs its server on 127.0.0.1, stores reviews in SQLite, and has no account and no telemetry. - It is built with a Rust core, Tauri and React. It runs on macOS and Linux. Windows is planned. - It is free and open source under Apache 2.0.
It is early, and I expect rough edges.
Site and downloads: https://pinrail.dev Docs: https://pinrail.dev/docs/
thanks for this - it's like the awesome plannotator... on a bit of steroids and, at least to my brain, cleaner.
[sorry, obligatory rotten emacs user who has to turn everything into a minor mode or at least configure emacs-like keyboard nav everywhere on my workstation]
Thanks for this.
I did some cursory poking around, and I'm glad to say, pinrail does not appear to be completely hopeless.
In particular,
- https://pinrail.dev/docs/building/settings-and-keys/#keyboar... suggests the possibility of potentially adding support for emacs-like keymap rebindings (e.g. <n>/<p> to move up and down, rather than <j>/<k>).
- https://pinrail.dev/docs/concepts/trust/#everything-is-on-yo... suggests that the system is meaningfully-pluggable and https://pinrail.dev/docs/building/design/#the-design-languag... suggests to me that that pluggability is not accidental.
Obviously, (a lot) more work is needed for even a starter emacs-flavored keymap or an emacs-native package to integrate with something like org-agenda.
But again, I'm glad that it does not (initially) appear to me that pinrail has painted itself into a corner in a way that would prevent that work from coming to fruition. And OP even seems amenable to FOSS contributions.
I might have to try it.